Integrative analysis with machine learning identifies diagnostic and prognostic signatures in neuroblastoma based on differentially DNA methylated enhancers between INSS stage 4 and 4S neuroblastoma

Ten cohorts were utilized in our study. Dataset GSE73518 containing transcriptome and methylation data of 105 NB patients was obtained from the GEO database. We utilized the champ.norm function of “ChAMP” R package to normalize the methylation data. Moreover, we obtained five independent transcriptome datasets including GSE49710 and GSE85047 cohorts from the GEO database, TARGET-NB cohort from the TARGET database, and E-MTAB-8248 and E-MTAB-179 cohorts from the ArrayExpress database. We removed patients with incomplete follow-up data and enrolled 1617 patients for the following analysis. We utilized the GSE49710 cohort to be the training cohort in which the prognostic and diagnostic models were established. Meanwhile, the other four datasets were set as the validation cohort to verify the stability of the models. Detail information on NB patients in bluk-seq cohorts was provided in Supplementary Table S1. The transcriptomic data was normalized with log2 (x + 1) algorithm and the combat function of “sva” R package was utilized to correct batch effects (Leek et al. 2012). Besides, three scRNA-seq datasets involving 31 patients with INSS 4 and 4S NB patients were obtained with accession number GSE192906, GSE137804, GSE140819 from GEO database. Another scRNA-seq cohort was downloaded from https://​www.​neuroblastomacel​latlas.​org/​, which contains scRNA-seq data published in a peer-reviewed article (namely CellAtlas dateset in the following analysis) (Kildisiute et al. 2021). Detail information on NB patients in four scRNA-seq cohorts was listed in Supplementary Table S2. The bioinformatics data used in this study is sourced from open-access platforms and is publicly available, thus negating the need for ethical approval or patient consent. The workflow diagram of our analysis is provided in Fig. 1.

Using the “ELMER” R package, we analyzed gene methylation and transcription data from NB samples in the GSE73518 cohort, focusing on the sequences of differentially methylated probes between INSS stage 4 and 4S NB. (Silva et al. 2019). Enhancer Linking by Methylation/Expression Relationships (ELMER) analysis revealed enriched motifs and predicted the transcription factors (TFs) interacting with these motifs, leading to the construction of a "TFs-motifs-genes" regulatory network. ELMER analysis involved five key steps: (1) Identifying distal probes (those more than 2 kb upstream from the transcription start site) in the methylation chip data; (2) detecting variations in methylation levels between normal and tumor samples; (3) determining target genes for the differentially methylated probes; (4) finding motifs enriched in probes related to both differential methylation and target genes; (5) identifying TFs based on transcriptional differences. From this, we obtained a list of 67 motif genes showing methylation differences between INSS stage 4 and 4S NB.

Based on the methylated differences between INSS stage 4 and 4S NB, we then exploited the INSS stage 4 and 4S related signature (ISRS) using motif genes enriched by ELMER analysis, to diagnose INSS stage 4 NB and predict the prognosis of NB patients. Totally, 67 different methylated genes in motif FOXK1_HUMAN.H11MO.0.A. were included in the construction of machine learning (ML) models. We integrated 10 ML algorithms involving random survival forest (RSF), elastic network (Enet), Lasso, Ridge, stepwise Cox, CoxBoost, partial least squares regression for Cox (plsRcox), supervised principal components (SuperPC), generalized boosted regression modeling (GBM) and survival support vector machine (survival-SVM) to predict prognosis. And we integrated 12 ML algorithms involving random forest (RF), Lasso, Ridge, elastic net (Enet), stepwise Glm, GlmBoost, LDA, partial least squares regression for Glm (plsRglm), GBM, XGB, SVM and Naive Bayes to diagnose stage 4 NB among all NB patients. Altogether 101 prognostic ML combinations and 113 prediction ML combinations were trained in the training cohort, to develop the prognostic and diagnostic models according to the leave-one-out cross-validation (LOOCV) framework. Models with < 5 genes were removed. The GSE49710 cohort was used as the training cohort, then the GSE85047, TARGET-NB, E-MTAB-8248 and E-MTAB-179 cohorts were used as the testing cohorts. Subsequently, the concordance index (C-index) of every ML combination in five cohorts was obtained (Liu et al. 2022). The top five ML combinations yielding the highest average C-index across five cohorts were chosen for further model evaluation via k-fold cross-validation, to mitigate overfitting and ensure the robustness and generalizability of the model. Area under the receiver operating characteristic curve (AUC), area under precision-recall curve (PRAUC), accuracy, sensitivity, specificity, precision, cross-entropy and Brier scores were calculated to identify the best diagnostic ML model via “mlr3” R package (Lang et al. 2019). Precision-recall curve (PRC) was employed to evaluate the performance of classification models in handling imbalanced datasets. Logarithmic loss, recall and decision calibration were utilized to select the best prognostic ML model via “mlr3proba” R package (Sonabend et al. 2021). We incorporated gene expression levels from various feature selection patterns to compute risk scores using a linear combination function for each ML combination, as dictated by the prognostic model. Similarly, we calculated the probability of stage 4 NB using the diagnostic model.

After selecting the most effective ML pattern pairs, we carried out thorough validation procedures to ensure ISRS with qualified accuracy, consistency, and reproducibility. In the prognostic signature, the median risk score from the training cohort was chosen as the threshold to categorize patients in both the training and validation cohorts into high or low-risk groups. We utilized the Kaplan–Meier (KM) survival analysis and the log-rank test on these groups, using the “survival” and “survminer” R packages. Cox proportional hazards model was utilized to identify the independent prognostic value of prognostic ISRS. In the diagnostic signature, we employed a confusion matrix to validate the precision of the signature via “cvms” R package. Receiver operating characteristic (ROC) curves, calibration curves and decision curve analysis (DCA) were employed to evaluate the precision, discrimination and clinical benefit of the diagnostic and prognostic signatures. Furthermore, we compared the performance of the prognostic signature against traditional clinical variables using time-dependent ROC curves. Additionally, both univariate and multivariate Cox analyses were conducted to affirm the independent predictive power of the prognostic signature.

Integrating the model genes from the diagnostic and prognostic signatures, we compiled a set of 24 genes which were used to construct our ML models. Utilizing these genes, we executed unsupervised clustering across three NB cohorts (GSE49710, E-MTAB-8248, and TARGET) using the “ConsensusClusterPlus” R package and the k-means algorithm (Wilkerson and Hayes 2010). This clustering was repeated 1000 times, each time including 80% of the samples. Following this, we visualized the heterogeneity between the two clusters using principal components analysis (PCA) and tSNE plots. To evaluate the efficacy of our clustering analysis, we compared the differences in clinicopathological features and gene expressions between patients of two subtypes using the “ComplexHeatmap” R package. Additionally, survival analysis was conducted to examine the outcomes between the different clusters.

Differentially expressed genes (DEGs) between two molecular subtypes based on consensus clustering analysis, as well as DEGs between two risk groups divided by ISRS, were screened via “limma” R package by False-discovery rate (FDR) < 0.05 and |log2fold change (FC)|> 1. To clarify the biological functions of DEGs between two subtypes and two risk groups, functional enrichment analysis was employed with Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) terms via “clusterprofiler” R package (Yu et al. 2012), and gene set variation analysis (GSVA) was performed with KEGG terms via “GSVA” R package (Hänzelmann et al. 2013). The “h.all.v7.4.symbols.gmt” hallmark gene sets from Molecular Signatures Database (MSigDB) were utilized for GSVA. Gene set enrichment analysis (GSEA) was conducted to identify molecular mechanisms and pathways associated with two subtypes and two risk groups (Subramanian et al. 2005), with the statistical significance threshold set as FDR < 0.25 and Normalized Enrichment Score (NES) > 1.

Initially, we applied eight distinct algorithms through “IOBR” R package, as well as single sample gene set enrichment analysis (ssGSEA), to measure the levels of immune infiltration in NB patients (Zeng et al. 2021; Newman et al. 2015; Yoshihara et al. 2013; Finotello et al. 2019; Li et al. 2016; Charoentong et al. 2017; Becht et al. 2016; Aran et al. 2017; Racle et al. 2017). Immune cell markers used in ssGSEA were identified from a literature review (Jia et al. 2018). We compared the differences of immune cell abundance between two risk groups divided by ISRS, and between two clusters divided by consensus clustering analysis, based on “Wilcox” test. Besides, we calculate the Spearman correlations between ISRS-predicted risk scores, expressions of model genes and immune cell contents quantified by various immune algorithms. Secondly, we utilized immune function markers identified from the literature review to conduct ssGSEA, assessing and comparing differences in immune function levels between two risk groups, based on “Wilcox” test (Barbie et al. 2009). Thirdly, we conducted ssGSEA to quantify the seven steps of the cancer immunity cycle based on gene markers from the Tracking Tumor Immunophenotype (TIP) website (http://​biocc.​hrbmu.​edu.​cn/​TIP/​) (Xu et al. 2018). Fourthly, we identified 24 inhibitory immune checkpoints from the literature review to compare differences in immune checkpoint gene expressions between two risk groups. Moreover, Thorsson et al. (2018) conducted immunogenomics analyses on over 10,000 cancer samples, identifying six immune subtypes that span a range of cancer subtypes and are believed to characterize immune response patterns, which potentially impact patient prognosis. Five immune subtypes were identified in the GSE49710 cohort, including wound healing (C1), IFN-gamma dominant (C2), inflammatory (C3), lymphocyte depleted (C4) and TGF-β dominant (C6). Subsequently, we focused on the distribution of each immune subtype in two risk groups and two clusters. To analyze the correlation between cancer-associated fibroblasts (CAFs) and immune cells, Estimate the Proportion of Immune and Cancer cells (EPIC), xCell, and Microenvironment Cell Populations-counter (MCP-counter) algorithms were utilized to obtain the CAF scores. And fractions of another 22 immune cells were assessed based on Cell-type Identification By Estimating Relative Subsets Of RNA Transcripts (CIBERSORT) algorithm.

Downloading the somatic mutation data of NB patients from the cBioPortal website (https://​www.​cbioportal.​org/​), we evaluated and visualized the mutation types and frequencies of the model genes via “maftools” R package (Mayakonda et al. 2018). In parallel, we calculated the tumor mutation burden (TMB) for each NB patient by determining the total number of somatic mutations per megabase (MB) in the exonic regions of the human genome. Gene mutations were categorized as either synonymous or nonsynonymous mutations. The nonsynonymous mutations involved Frame_Shift_Del, Frame_Shift_Ins, In_Frame_Del, In_Frame_Ins, Missense, Nonsense, Nonstop, Splice_Site, and Translation_Start_Site. We employed GISTIC 2.0 to identify key mutation regions based on copy number variation (CNV) data sourced from the cBioPortal website (Mermel et al. 2011). Additionally, the frequency of somatic CNVs in model genes among NB patients was graphically represented using “bubble plots”, and the chromosomal locations of these mutations were depicted through “circle plots” using the “RCircos” R package (Zhang et al. 2013).

To ascertain the effectiveness of ISRS in predicting responses to immunotherapy, we calculated the immune dysfunction and exclusion (TIDE, http://​tide.​dfci.​harvard.​edu/​) score in two risk groups. Besides, the submap algorithm was employed to assess the response sensitivity of immunotherapy in two risk groups based on a public dataset of immunotherapy (Jiang et al. 2018; Roh et al. 2017). Moreover, four immunotherapy-treated cohorts, IMvigor210, GSE78220, GSE135222, and GSE91061, were obtained to appraise the efficacy of ISRS in predicting responses to immunotherapy. Subsequently, we collected the chemotherapy sensitivity of human cancer cell lines to various drugs from the Cancer Therapeutics Response Portal (CTRP, https://​portals.​broadinstitute.​org/​ctrp) website and Profiling Relative Inhibition Simultaneously in Mixtures (PRISM, https://​depmap.​org/​portal/​prism/​) website. The cell line that is more sensitive to a potential drug would display a lower area under the curve (AUC), which might help screening novel drugs for high-risk NB patients identified by ISRS (Yang et al. 2021).

The single-cell RNA sequencing (scRNA-seq) data obtained from the GEO database with accession number GSE137804 was created as Seurat objects via “Seurat” R package (4.1.0) (Satija et al. 2015). We performed quality control to exclude low-quality cells with unique feature counts > 6000 or < 300 or mitochondrial counts > 15%, as well as ribosomes < 3% and erythrocytes < 0.1%, resulting in 172,564 cells finally. We utilized Harmony R package to mitigate technical batch effects while retaining biological variation during multiple batch integration (Korsunsky et al. 2019). The FindVariableFeatures function was performed to identify the top 2000 genes with the highest variation between cells (Butler et al. 2018), on which PCA was performed in the expression matrix. We set resolution at 0.3 to conduct the FindClusters function to identify various clusters. To ascertain the predominant cell type expressing the model genes of ISRS, we performed the RunUMAP function in 4 scRNA-seq datasets (GSE192906, GSE137804, GSE140819 and CellAtlas). We manually annotated the cell type of every cluster based on the expression of canonical markers from the literature review (Dong et al. 2020). Six scoring algorithms (AUCell in “AUCell” R package, Ucell in “Ucell” R package, ssGSEA in “GSVA” R package, singscore in “singscore” R package, AddModuleScore and PercentageFeatureSet in “Seurat” R package) were utilized to perform signature enrichment scoring in four NB scRNA-seq datasets (Hänzelmann et al. 2013; Satija et al. 2015; Andreatta and Carmona 2021; Foroutan et al. 2018; Aibar et al. 2017), which were visualized with “irGSEA” R package (https://​chuiqin.​github.​io/​irGSEA/​index.​html). Pseudotime trajectory analysis was conducted through “Monocle” R package and “Monocle3” R package to explore map conversion trajectories without prior knowledge of differentiation time or direction (Qiu et al. 2017; Cao et al. 2019). “InferCNV” R package was utilized to assess CNVs in neuroendocrine (NE) cells, Schwann cells, endothelial cells, and fibroblasts. T cells, B cells and myeloid cells were considered as references to assess CNVs in cancerous cells (Tirosh et al. 2016). “CellChat” R package was used to explore the intercellular communication patterns between each cell type (Jin et al. 2021). Meanwhile, “pySCENIC” (version 0.11.2) in Python (version 3.7) was performed to investigate TF enrichment and regulon activity, which aided in the construction of TF regulatory networks and the identification of stable cell states (Aibar et al. 2017).

We performed pan-cancer analysis via “TCGAplot” R package to explore hub genes’ similarities and differences in genomic and cellular changes across a variety of tumor types (Liao and Wang 2023), in terms of gene expression, TMB, microsatellite instability (MSI), immune microenvironment and prognostic value. Gene expressions in tumor and normal samples were compared by “Wilcox” test, while the correlation between the gene expression and TMB, MSI, immune cell and immune score was calculated by the Spearman method. Immune cell ratio was retrieved from The Immune Landscape of Cancer (https://​api.​gdc.​cancer.​gov/​data/​b3df502e-3594-46ef-9f94-d041a20a0b9a), and immune scores were analyzed by the ESTIMATE algorithm.

To validate the different expressions of model genes between stage 4 and 4S NB, we conducted immunohistochemistry (IHC) staining of two model genes (CAMTA2 and FOXD1) in 25 stage 4 NB tissues and 10 stage 4S NB tissues. The experiment received approval from the ethics committee of the Children’s Hospital of Chongqing Medical University. NB tissues were paraffin-embedded and cut into 4 mm slices. Following dewaxing, hydration, and antigen retrieval, the samples were treated with primary antibodies: Anti-CAMTA2 (Affinity Biosciences Cat# DF9314, RRID: AB_2842510) and Anti-FOXD1 (bs-12193R, Bioss, Beijing, China), and incubated overnight at 4 °C. Subsequent steps included incubation with Goat anti-Rabbit IgG secondary antibody (ZENBIO, China), DAB staining (ZENBIO, China), and blocking, with the staining effects observed under a microscope. Each sample was scored based on staining intensity (0: none, 1: mild, 2: moderate, 3: strong) and the percentage of positive cells (0: 0%, 1: 1–25%, 2: 26–50%, 3: 51–75%, 4: 76–100%). The final IHC score was calculated as the sum of both intensity and percentage scores.

Methylation and transcriptomic data of the same patient, particularly in 17 INSS stage 4 NB patients (age < 1.5 years) and 20 INSS stage 4S NB patients from the GSE73518 cohort, were incorporated in ELMER analysis. We plotted a heatmap to visualize the variations of methylation and transcriptomic levels between stage 4 and 4S NB (Fig. 2A). Distal probes were used to detect the region of enhancers. According to the hg38 reference genome, we selected 135,427 distal probes by get.feature.probe function (Supplementary Table S3). We utilized the GetNearGenes function to pinpoint the top ten genes nearest to the upstream and downstream of the distal probes respectively, thus creating probe-gene pairs (Fig. 2B). Subsequently, an unsupervised approach was utilized to detect distal probes by get.diff.meth function. Each probe’s methylation levels were ranked in both stage 4 and 4S groups, and the samples in the lower or higher quintiles (20%) of methylation were analyzed to determine whether the probes were hypo- or hypermethylated in stage 4, leading to the identification of 57 distal probes (Supplementary Table S4) with an FDR < 0.01 and a |Δβ| < 0.3 (Fig. 2C). We then examined the inverse correlation between each probe's methylation level and its associated gene's expression. Samples were classified into methylated (M) and unmethylated (U) groups based on the top and bottom 20% methylation levels of the probes, and gene expression levels between these groups were compared using the Mann–Whitney U test. Unsupervised mode was utilized to screen 956 probe-gene pairs with statistically significant negative correlations with default parameters through get.pair function (Supplementary Table S5). Then, the 250 bp base sequence upstream and downstream of these probes were extracted and aligned to 37 significantly enriched motifs (Supplementary Table S6) by get.enriched.motif function. Distal probes associated with the same motif were then classified into the top 20% M and bottom 20% U groups based on methylation levels. Motif FOXK1_HUMAN.H11MO.0.A. with differential expressions between two stages and negative correlations with its methylation levels was obtained through get.TFs function via the unsupervised mode (FDR < 0.05) (Fig. 2D–F, Supplementary Table S7) (Lambert et al. 2018). To improve the robustness of the identified regulatory networks, we also employed a supervised mode for probe selection and motif identification, which revealed that motif FOXK1_HUMAN.H11MO.0.A. enriched differentially between two stages, with significance (FDR < 0.05) (Supplementary Table S8). Ranking the TFs in motif FOXK1_HUMAN.H11MO.0.A. discovered that EPAS1 and L3MBTL4 may play a vital role in methylated regulation (Fig. 2G). Besides, most of the genes in motif FOXK1_HUMAN.H11MO.0.A. expressed significantly higher in stage 4S than in stage 4 (Fig. 2H). GO and KEGG enrichment analysis showed that motif genes may affect embryonic organ development, muscle tissue development, morphogenesis and signaling pathways regulating pluripotency of stem cells (Fig. 2I, J). The enhancer-associated regulatory network is visualized via Cytoscape software (Fig. 2K).

As mentioned above, five NB cohorts with transcriptome data were utilized in model development and verification, which showed a superior effectiveness of batch effect removal via “sva” R package (Supplementary Figure S1A–B). Incorporating 67 genes in motif FOXK1_HUMAN.H11MO.0.A. based on ELMER analysis, 101 prognostic algorithm combinations and 113 prediction algorithm combinations were then constructed via the LOOCV framework. The C-index of each ML combination was calculated in all validation datasets (Supplementary Table S9). The best diagnostic model was established based on RF algorithm which was utilized in both feature selection and model construction, with the highest average C-index (0.812) across five datasets (Fig. 3A). AUC, PRAUC, accuracy, sensitivity, specificity, precision, cross-entropy and Brier scores were utilized to reveal that RF was the most powerful model in diagnosing INSS 4 NB (Supplementary Figure S1C–D). Finally, a 9-gene diagnostic INSS stage-related signature (ISRS) was accordingly established to diagnose stage 4 NB, with CAMTA2 being the most important variable (Supplementary Figure S1E–F). Confusion matrix in E-MTAB 8248 cohort and other three validation cohorts showed a well accuracy of ISRS (Fig. 3B, Supplementary Figure S2A). ROC curves of five cohorts showed a well discrimination of ISRS (Fig. 3C). We also compared the AUC of ISRS, other clinical variables and the logistic regression nomogram model including several clinical variables and ISRS, which demonstrated that ISRS and the nomogram model performed better (Fig. 3D, Supplementary Figure S2B–C). Calibration curves showed a great alignment between ISRS predicted probability and the observed probability of stage 4 NB (Fig. 3E). DCA curves indicated that utilization of ISRS and the logistic regression nomogram model is more clinically beneficial, compared with other clinical variables (Fig. 3F, Supplementary Figure S2D). The feature importance visualization of 9 variables selected by RF demonstrated that CAMTA2 impacted most in the RF model (Fig. 3G). The best prognostic model was established based on the Enet (alpha = 0.6) algorithm which was utilized in both feature selection and model construction, with the highest average C-index (0.732) across five datasets (Fig. 3H). Logarithmic loss, recall and decision calibration were calculated to prove the well calibration and precision of the Enet (alpha = 0.6) model (Supplementary Figure S1G). Ultimately, a 20-gene prognostic ISRS was accordingly established to forecast the prognosis of NB patients (Supplementary Figure S1H–I). The feature importance visualization of 20 variables selected by Enet demonstrated that CAMTA1 impacted most in the Enet model (Supplementary Figure S1J). In GSE49710, E-MTAB 8248, TARGET and E-MTAB 179 cohorts, the low-risk group owned a relatively longer overall survival (OS) and event-free survival (EFS) than the high-risk group (Fig. 3I, Supplementary Figure S3A). In GSE85047 cohort, the low-risk group owned a relatively longer overall survival (OS) and progression-free survival (PFS) than the high-risk group (Supplementary Figure S3A). ROC curves of 1-, 3- and 5-year OS showed well specificity of ISRS (Fig. 3J, Supplementary Figure S3B). AUC values of 3-year OS proved that ISRS and cox regression model involving ISRS and several clinical variables were more specific and discriminative to forecast the prognosis of NB patients, compared to other clinical variables (Fig. 3K, Supplementary Figure S3C–D). Time dependent ROC curves indicated that ISRS and cox regression model outperformed conventional clinical variables in capability of discrimination (Fig. 3L, Supplementary Figure S3E). Calibration curves (Fig. 3M, Supplementary Figure S3F) and DCA curves (Fig. 3N, Supplementary Figure S3G) showed that ISRS is well-behaved in accuracy and clinical benefit. Multivariate Cox regression analysis in the Cox proportional hazards model indicated that ISRS, age, sex, INSS stage and COG risk stratification were independent prognostic factors in NB patients (P < 0.05) (Fig. 3O, Supplementary Figure S3H). All these metrics collectively indicated that ISRS demonstrated stability and robustness in model performances across five NB queues. The flowchart of machine learning algorithm integration was provided in Supplementary Figure S1K.

PCA analysis displayed significant variations between two risk groups divided by prognostic ISRS (Fig. 4A). Based on 7325 DEGs identified between two risk groups via “limma” R package, we performed functional enrichment analysis according to GO and KEGG terms, which discovered that DEGs were significantly enriched in dopamine secretion, kinetochore organization, serotonin transport and outer kinetochore in GO terms, and were significantly enriched in Neuroactive ligand-receptor interaction, Cell cycle, GABAergic synapse and Calcium signaling pathway in KEGG terms (Fig. 4B, C). GSEA analysis demonstrated that carbon metabolism and cell cycle were enriched in a high-risk group, and adrenergic signaling in cardiomyocytes, aldosterone synthesis and secretion and cell adhesion molecules were suppressed in a high-risk group (Fig. 4D, E). To understand the biological behavioral variations between two risk groups, we conducted GSVA enrichment analysis based on “h.all.v7.4.symbols.gmt” hallmark gene sets in MSigDB database, which indicated that high-risk group was enriched in myc_targets_v2 and DNA_repair, and suppressed in hedgehog_signaling and apical_surface (Fig. 4F). We then visualized the differential expression of ISRS model genes and the variations of clinical variables between the two risk groups (Fig. 4G). ZEB2, CAMTA1, BAZ2B, CAMTA2 and HOXC9, which constituted both the diagnostic and prognostic ISRS, were significantly higher expressed in a low-risk group than in a high-risk group, and significantly higher expressed in stage 4S than in stage 4, pretending to be protectively prognostic genes (Supplementary Figure S4A-B). Detailed information on types of model genes was provided in Supplementary Table S7. Spearman correlation analysis revealed close relationships (correlation P value < 0.0001) among the 24 ISRS model genes (Fig. 4H). Based on CNV data sourced from the cbioportal website, “bubble plots” visualized that CREB5 attained the most somatic CNV frequencies among diagnostic model genes, and CAMTA1 attained the most somatic CNV frequencies among prognostic model genes (Fig. 4I, J). Besides, “circle plots” visualized the chromosome locations where the mutations of ISRS model genes occurred (Fig. 4K).

To appraise the discriminative capability of the prognostic ISRS in immune infiltration, we simultaneously calculated the abundance of immune cell infiltration based on eight distinct immune algorithms. “ComplexHeatmap” R package visualized that various immune cells were significantly fewer infiltrated in a high-risk group than in the low-risk group (Fig. 5A). Besides, the spearman correlation heatmap displayed the correlation between immune cell infiltrations and the risk scores calculated by the prognostic ISRS, as well as the relationship between immune cell infiltrations and the expression of ISRS model genes (Fig. 5B). Besides, ssGSEA analysis of immune function scores demonstrated that the low-risk group owned significantly better immune infiltration abundance (Fig. 5C). Moreover, the activation of six key steps in the cancer immunity cycle appeared to be significantly higher in the low-risk group (Fig. 5D). For immune checkpoints, the low-risk group displayed elevated expression levels of immune checkpoint genes, indicating a susceptibility to immunotherapy (Fig. 5E). Previous literatures have stressed the vital role of cancer-associated fibroblasts (CAFs) in immune modulation of the tumor microenvironment (Boyle et al. 2020; Sahai et al. 2020; Gagliano et al. 2020). We then utilized “circle plot” to visualized the correlation of CAFs among various types of immune cells calculated by the CIBERSORT algorithm in NB patients (Fig. 5F). Additionally, it is well-known that heterogeneous metabolic preferences and dependencies exist across tumor types (Hakimi et al. 2016; Hensley et al. 2016; Kim and DeBerardinis 2019), hence we extracted multiple metabolic pathways from the KEGG database to validate the relationship between risk scores and tumor metabolism in NB patients (Fig. 5G). Two hub model genes (FOXD1 and CAMTA2) also showed significant correlations with cell metabolic pathways.

We visualized and compared the distribution variations of somatic mutations between the low-risk group (Fig. 6A) and the high-risk group (Fig. 6B) based on mutation data sourced from the cbioportal website. Comparisons of tumor mutation burden (TMB) between the two risk groups revealed no significant difference (Figs. 6C). However, TMB was found to be significantly related to the risk scores obtained by the prognostic ISRS based on spearman correlation analysis (Fig. 6D). Moreover, we classified NB patients with mutation data into the high TMB and the low TMB group according to their median TMB score. After integrating two risk groups and two TMB groups, we found that patients with low TMB in the high-risk group owned the worst OS and EFS, and patients with high TMB in the low-risk group owned the best OS and EFS, without significance (Fig. 6E, F).

Based on TIDE scores and submap algorithm, the possibility of immunotherapy responses was appraised in two risk groups. In the GSE49710 cohort, higher TIDE scores, higher TIDE dysfunction and exclusion scores were discovered in the high-risk group, indicating a bigger probability of immune escape during the immunotherapy process (Fig. 6G). In the IMvigor210 immunotherapy cohort, survival analysis demonstrated that low-risk patients with a response to immunotherapy owned significantly longest OS, whereas high-risk patients with no response to immunotherapy had significantly worst OS (Fig. 6H). Meanwhile, in the E-MTAB8248 cohort, a significantly higher proportion of patients responding to immunotherapy was observed in the low-risk group (Fig. 6I). Besides, the submap analysis results demonstrated that low-risk patients were more sensitive to CTLA4 inhibitors, rather than PD1 inhibitors (Fig. 6J). Despite we assessed individual immunotherapy reaction based on two methods in two NB cohorts, it's crucial to straightly compare treatment effectiveness in immunotherapy-treated cohorts between two risk groups, which called a need to involve four immunotherapy-treated cohorts in subsequent analysis. Ultimately, we found that patients who responded more sensitively to immunotherapy owned significantly lower risk scores across four immunotherapy-treated cohorts (Figs. 6K). Accordingly, to explore potential drugs that might be more effective in high-risk NB patients, we predicted drug response based on drug sensitivity data from CTRP and PRISM. Through the cross-correlation in the two pharmacogenomics databases, we successfully predicted six potential drugs or compounds (BI-2536, daporinad, GSK461364, SB-743921, rubitecan and talazoparib) with therapeutic effects in high-risk NB patients (Fig. 6L).

To further understand the expression pattern of ISRS model genes, 871 patients from GSE49710, E-MTAB 8248 and TARGET cohorts were included for consensus clustering analysis. Respectively, we utilized ISRS model genes to employ unsupervised clustering analysis on each cohort, which demonstrated that k = 2 exhibited the best discrimination (Supplementary Figure S5A–B). Besides, PCA and tSNE analysis showed significant variations between the two clusters, according to the expressions of ISRS model genes (Fig. 7A, Supplementary Figure S5C). Meanwhile, we performed survival analysis to demonstrate that cluster 1 owned significantly worse OS and EFS than cluster 2 (Fig. 7B, C). Moreover, the expressions of ISRS model genes and the clinicopathological characteristics in different clusters displayed significant differences (Fig. 7D). Based on 1868 DEGs identified between two clusters via “limma” R package, we performed functional enrichment analysis, which discovered that DEGs were significantly enriched in neuropeptide signaling pathway, neuropeptide hormone activity, ligand-gated ion channel activity, sodium:chloride symporter activity and GABA receptor activity in GO terms, and were significantly enriched in Neuroactive ligand-receptor interaction, ECM-receptor interaction, GABAergic synapse, Synaptic vesicle cycle and Neomycin, kanamycin and gentamicin biosynthesis in KEGG terms (Fig. 7E, F). GSEA analysis demonstrated that cytokine-cytokine receptor interaction and staphylococcus aureus infection were enriched in cluster 1, and GABAergic synapse and neuroactive ligand-receptor interaction were suppressed in cluster 1 (Supplementary Figure S5D). Moreover, GSVA enrichment analysis, based on “h.all.v7.4.symbols.gmt” hallmark gene sets in the MSigDB database, indicated that cluster 1 was enriched in myc_targets_v2, e2f_targets and g2m_checkpoint, and suppressed in hedgehog_signaling and apical_junction (Supplementary Figure S5E). To further understand variations in tumor immune microenvironment between two clusters, eight immune algorithms were utilized to assess the immune abundance differences between two clusters, and visualized the Cox P value of each immune cell type (Fig. 7G). In terms of mutation landscape, we compared the distribution variations of somatic mutations between cluster 1 (Supplementary Figure S5F) and cluster 2 (Supplementary Figure S5G) based on mutation data sourced from the cbioportal website. We observed that TMB scores are not significantly different between the two clusters (Supplementary Figure S5H). Meanwhile, we classified NB patients with mutation data into the high TMB and the low TMB group according to the median TMB score. After integrating two clusters and two TMB groups, we found that patients with low TMB from cluster 1 had worse OS and EFS than patients with high TMB from cluster 2, without significance (Supplementary Figure S5I–J).

For the sake of verifying the prognostic effectiveness of ISRS, we collected coefficients of model genes in 39 previously published NB prognostic models. Subsequently, we performed a comparative analysis of the C-index of each prognostic model, which was developed based on a variety of biological characteristics, involving necroptosis, ferroptosis, cuproptosis, disulfidptosis, ganglioside, and m6A methylation. Ultimately, we discovered that ISRS demonstrated superior predictive performances relative to the vast majority of models across five NB datasets (Fig. 8A), which qualified ISRS as an invaluable NB prognostic model. Moreover, the overall relationship among different risk groups, different clusters and patients’ clinical information was visualized by “sankey plot” (Fig. 8B). In terms of immune subtypes of NB patients, we observed significant variations in the proportion of immune subtypes between two risk groups and two clusters in the GSE49710 cohort, with much more wound healing (C1) subtype in high-risk group and in cluster 1 (Fig. 8C). Comparing clinical variables between two risk-groups, we suggested that lower risk scores of ISRS were related to better prognoses in NB patients (Fig. 8D).

To explore the biological function of ISRS in a single cell level, we utilized 4 NB single-cell datasets (GSE137804, GSE192906, GSE140819 and CellAtlas) to validate our results. In GSE137804, we utilized harmony integration to remove batch effects, which showed a well correction before and after integration (Fig. 9A). Subsequently, we partitioned all cells in GSE137804 into 25 clusters based on the k-nearest neighbor (KNN) clustering method (Fig. 9B). According to cell markers collected from the literature review, we successfully identified 8 distinct cellular subtypes, involving neuroendocrine cells (NE cells), T cells, B cells, endothelial cells, myeloid cells, Schwann cells, fibroblasts, and plasmacytoid DC cells (Fig. 9C). Violin plot visualized the expressions of canonical markers in 8 clusters (Fig. 9D). Subsequently, the ISRS scoring for each cell was calculated based on six signature enrichment scoring algorithms, which demonstrated that cells with higher ISRS scoring predominately located in NE cells (Fig. 9E, F, Supplementary Figure S6A–B). Meanwhile, we explored the single-cell transcriptome localization of ISRS model genes, which were found to expressed mostly in NE cells (Fig. 9G). Accordingly, the expression patterns of ISRS model genes were validated in three NB single-cell external datasets (GSE192906, GSE140819 and CellAtlas) (Supplementary Figure S6C). Meanwhile, six scoring algorithms were utilized in the three NB single-cell external datasets (GSE192906, GSE140819 and CellAtlas) to demonstrate that ISRS model genes were enriched in NE cells (Supplementary Figure S6D–I). Additionally, the pseudotime trajectory analysis revealed the temporal sequence of malignant cellular differentiation and demonstrated that NE cells with higher ISRS scores appeared at an earlier pseudotime than NE cells with lower ISRS scores, indicating that immature NE cells were more predominant in high-risk NB patients (Fig. 9H). The differentiation trajectory of NE cells, endothelial cells, fibroblasts and Schwann cells were plotted via Monocle 2 algorithm, which suggested that endothelial cells owned the potential to differentiate into other types of cells (Fig. 9I). Hence, we set endothelial cells as the beginning of differentiation to investigate pseudotime trajectory of all cells in Monocle 3 analysis, which validated that immature NE cells could own higher ISRS scores and serve as malignant cells (Fig. 9J). Besides, pseudotime analysis of interested genes identified that pseudotime DEGs were increased along the pseudotime curve, while the ISRS model genes stayed still (Fig. 9K).

To comprehensively verify the efficacy of ISRS in single-cell RNA-sequencing data, we utilized inferCNV algorithm to investigate the clonal structure of NE cells, endothelial cells, fibroblasts and Schwann cells. The infer CNV analysis revealed that NEs with chromosome 17q gain could serve as malignant cells, while cells with more CNVs owned higher ISRS scores in 4 NB single-cell datasets, validating the capability of risk stratification of ISRS (Fig. 10A, B, Supplementary Figure S7A–C). In cell communication analysis, we plotted circle diagrams to demonstrate the interaction times and interaction strengths among each cell type, visualizing the integrated cell communication networks in high ISRS cells and low ISRS cells. We compared the cell communication patterns between two ISRS groups, which revealed that B cells, myeloid cells, Schwann cells and fibroblasts might contribute most to differences in cell communication networks (Fig. 10C). Hence, we investigated over-expressed ligand-receptor pairs and their interactions among B cells, myeloid cells, Schwann cells and fibroblasts in the high ISRS group, identifying differential cell interactions between two ISRS groups (Fig. 10D). Different cell types would emit different contributive signals on the overall, incoming and outgoing signals between two ISRS groups, especially for Schwann cells, fibroblasts and endothelial cells (Fig. 10E, F, Supplementary Figure S7D). Besides we further investigated the relation between regulon (TFs and target genes) activity and each cell type in different ISRS group based on SCENIC analysis, which indicated that regulons of SOX4 and SMAD5 scored high activity in high ISRS cells, and regulons of KLF7 and SMARCA4 scored high activity in low ISRS cells (Fig. 10G, Supplementary Figure S7E).

Finally, we observed two ISRS model genes (CAMTA2 and FOXD1) with abundant research value in oncology, which have been proven as oncogenes in head and neck squamous cell carcinoma (Wu et al. 2023) or colon cancer (Luan et al. 2021). Hence, we performed pan-cancer analysis to reveal the heterogeneity of CAMTA2 and FOXD1 expressions in tumor and normal samples across 33 tumor types (Fig. 11A). Meanwhile, correlation between expressions of two hub genes and TMB, MSI, immune cell and immune score highlighted the significance of two hub genes involved in tumor immune microenvironment and immune cell infiltration (Fig. 11B–E). The protective prognosis value of CAMTA2 was observed in PAAD, while that of FOXD1 was found in CESC and LUSC, showing a similar prognostic effect in NB (Fig. 11F). Furthermore, we conducted IHC staining to validate the significantly higher protein level of CAMTA2 and FOXD1 expression in stage 4S tissues than in stage 4 tissues, which supported our bioinformatics results (Fig. 11G, H, Supplementary Figure S7F). Meanwhile, we divided 35 stage 4S and 4 NB patients into high and low CAMTA2/FOXD1 group based on the median IHC score of CAMTA2/FOXD1. Survival analysis revealed that NB patients with high CAMTA2 protein levels had significantly better OS than NB patients with low CAMTA2 protein levels, while this difference was also observed between high and low FOXD1 groups, without significance (Fig. 11H).